Fig. 8

Interpolation of gene expression patterns is accurate. a Scatterplot of classifiers trained on gene expression profiles of one condition, which are subsequently used to predict labels of either measured expression profiles from the same condition in a cross-validation framework (x-axis) or used to predict labels of cells sequenced from the other condition that were then interpolated into this condition (y-axis). Similarity in accuracy represented by points near the diagonal indicates that cell type identity encoded in the gene expression profile is maintained even after interpolation. b The pairwise cell-cell similarity matrix for all cells projected into the young condition, including both the old cells interpolated into the young condition (yellow) and the cells originally sequenced in the young condition (blue). Note that cells cluster largely by cell type regardless of the condition in which they were sequenced. c The pairwise cell-cell similarity matrix for all cells computed using the original expression measurements. d The pairwise cell-cell similarity matrix for all cells computed using the low-dimensional coordinates within the alignment space learned by scAlign. Similarity between c and d indicates the scAlign embedding maintains global similarity patterns between cells in the original gene expression space